fireviewer/rfdetr-medium-probable-fire-v2.5

RF-DETR Medium · probable fire · v2.5

Quick card

Field	Reference
Use	Generate single-class probable_fire candidate regions from full images.
Expected input	Full RGB image resized to 1088 px; exact RF-DETR preprocessing.
Example	An image → scored xyxy candidate boxes in original-image coordinates.
Limits	Flame and smoke are merged at this proposal stage. Selected-validation scores are not the external cascade TEST.
Contract	English pipeline · Reproducible loading

Choose a variant · Resource navigation · Update log

purpose and evidence

Single-class full-image proposal detector: probable_fire merges visible flame/fire and smoke. It supplies candidate regions to the visible-flame and optional smoke branches; it does not distinguish them itself. Input is RGB resized by RF-DETR to 1088 px; output is scored xyxy boxes in original-image coordinates. Requested 1080 was aligned to 1088; this release is not a 1080 model.

RF-DETR Medium 1.11.0, positional encoding size 68, seed 42, batch 16, 15 completed epochs, cosine learning-rate schedule with one warmup epoch. Selected checkpoint is epoch 10, validated from its payload and selection record, not the stale end-of-run filename index. Selection criterion: 0.5 AP-small + 0.5 small-object recall at score 0.25 / IoU 0.50. Training used 60,891 images and validation 8,288 from the retained deduplicated corpus. The linked public v2.5-box-precise-and-multi-source-suplement is the rights-filtered subset (46,576 train / 7,647 validation / 9,838 test), not the complete training corpus.

Selected validation result

Full retained validation: 8,288 images, 6,898 probable-fire boxes. Fixed score 0.25 and greedy IoU 0.50 matching for precision/recall; this score is a comparison threshold, not a calibrated production alert threshold. COCO small uses original-image box area ≤32² px. Selection metrics are validation measurements, not an independent TEST benchmark.

Metric	Value
COCO mAP50–95	0.419858
AP50 / AP75	0.702712 / 0.424334
AP-small / AR100	0.275894 / 0.638069
Precision / Recall / F1 @0.25	0.470818 / 0.785880 / 0.588855
TP / FP / FN	5,421 / 6,093 / 1,477
Small recall (943 / 1,357 GT)	0.694915
Negative-image false alert rate (842 / 4,774)	0.176372

The negative-image statistic is the fraction of annotated-negative images with at least one surviving detection, not a per-box FP rate. AP and operating-point counts are different measures. Missing source labels can penalize real detections.

Checkpoint identity

• Run: 20261004-202150-rfdetr-ccdb, selected E10, completed E15.
• Original checkpoint SHA-256: 05e431638e3daf9c81c78860ad26b8b65c52765b9d934666d9d4d7da66eb565d.
• Public tensor SHA-256: 3157e4d09f90061b0a8deb9fd42b2db60d5978fb55872d642c5aa75338377a26.
• Training-corpus fingerprint: ab4360d4ef3f6e1f419a317b9d8ddff5602c8e3e3394babd1bd0f3faa8916efe.

The public tensor package is value-identical to the selected checkpoint tensors, finite, and excludes optimizer/RNG state. Full continuation checkpoints and the requested epoch 8 backup remain private on Drive. The different file hashes reflect different containers.

Reproducible loading

The release contains tensor-only model.safetensors, a plain config.json, an explicit inference.py and checksum provenance. Download a revision-pinned snapshot, install the repository's requirements, and load it explicitly. No Hugging Face trust_remote_code is required. The optional smoke backbone fetches a pinned official Facebook Research Torch Hub source; inspect and cache that source for an offline deployment. RF-DETR 1.11 requires a native wrapper; the helper creates it locally from safe tensors and JSON, never from an external pickle.

from huggingface_hub import snapshot_download
import importlib.util
root = snapshot_download(REPOSITORY, revision=REVISION)
spec = importlib.util.spec_from_file_location('fireviewer_inference', root + '/inference.py')
fv = importlib.util.module_from_spec(spec)
spec.loader.exec_module(fv)
model, config = fv.load_model(root, device='cpu')  # or cuda

Set REPOSITORY to this repository ID and REVISION to a full commit SHA from its Files tab. The complete inference CLI and geometry are documented in PIPELINE_EN.md and PIPELINE_FR.md. This is an explicit PyTorch research interface, not a Transformers AutoModel export.

Scope, rights and limitations

These checkpoints are retained research results. The original TEST set was locked during selection; no final held-out TEST score is claimed. Exact-byte/decoded-image duplicate checks were performed upstream, but event-level independence and deployment-domain transfer remain unestablished. Public linked data are rights-filtered subsets of the retained training corpus, not a byte-for-byte reproduction of the full training set. Private MyData_Fire and undocumented With-Augmentation-Big-1 images are absent from public datasets. Publicly releasing a checkpoint does not establish deployment qualification.

The model/code release uses Apache-2.0; RF-DETR and DINOv2 upstream notices must be retained. Image rights are separate and source-specific. Predictions are visual estimates, with possible missed fire, smoke confusion, lighting artifacts and source annotation omissions. They are not a calibrated emergency alarm or geographic ignition estimate. See each linked dataset's notices before reusing images.

Official RF-DETR source and citation.

LiteRT FP32 export

The selected weights above are also available as a verified LiteRT inference package. It includes real CPU PyTorch/LiteRT parity evidence, output/class/preprocessing contracts, an Android ModelConfig and a Cadryl StudioPack. Static batch 1; no new training or quantization; Android-device qualification remains open. Full English/French cascade and external benchmark.

Likes are always welcome! ❤️ If these resources help you, a little heart makes our day and helps others discover the work. Thank you!